Three examples of counter-AI technologies are neural networks, blockchain-based digital signatures, and machine-speed sensors.
- A neural network utilizes machine learning to process various details and technical aspects of artworks, images, or videos.
- The complete analysis is used to identify visual patterns.
- Such algorithms can also "detect synthetic currency notes, historical evidence, and documents."
- Additionally, they are able to compare specific videos and images across multiple sources to find out the original one.
- If those algorithms are properly "trained," they also recognize anomalies, such as the untypical movements of a person in a video.
- The researchers from UC Berkeley developed a neural network that can detect deepfake videos based on unnatural head movements and facial expressions with 92% accuracy. Their research can be found here.
- The downside of the method is that the algorithm has to be familiar with the typical movements of each individual. It can be trained to analyze deepfake video news involving politicians and celebrities, but it is less efficient for news about regular people.
- Also, the Video Computing Group at the University of California used a deep neural network architecture for detecting fake images at a pixel level. As the researchers explained, if the image has been tampered with, pixels around objects that have been either removed or added share certain characteristics, like unnatural smoothness. While they won't be recognized by a human eye, a deep learning algorithm is able to detect them.
- Their neural network gathered patterns that describe the difference between non-manipulated and manipulated images. More details about the research can be found here.
- A San Diego-based company, TruePic, uses deep learning as part of their proprietary Controlled Capture technology. Their tool tests all photos and videos, comparing them against various sets of data, to establish authenticity.
- Blockchain is a distributed ledger that allows storing information online, as opposed to a centralized server. In this technology, "every record is replicated on multiple computers and tied to a pair of public and private encryption keys."
- The person who holds the private key truly owns the data.
- At the moment, the solution is not very efficient for large data sets. However, it works well for digital signatures.
- To counter deepfake news articles, videos, and images, people can sign them using digital signatures.
- The more people sign, the more likely it is that the video is original. The downside is that it remains to be decided how to select people who are permitted to sign.
- One company that uses digital signatures in their verification platform is TruePic. Each photo and video taken using their patented technology is signed using a blockchain-based solution. It is one of the measures taken to ensure authenticity and protect their clients' data, along with using deep learning and video forensics.
- Also, a non-profit organization, PUBLIQ, encrypts all of its publications using blockchain-based technology. Afterward, they "are distributed to a peer-to-peer network," which makes them more trustworthy.
- Machine-speed sensors are a technology that often compliments blockchain-based digital signatures, eliminating doubts about whether a digital signature was added legitimately.
- It involves the use of data provided by device sensors, which is irrefutable and captured in real-time.
- The kind of data recorded can include the date, time, GPS location, altitude, as well as mobile and WiFi networks in the area.
- An app that records sensor data to prove the credibility of photos and videos is ProofMode. It works in the background of the device all the time, and captures metadata each time a person takes a photo.
- ProofMode is a tool for journalists and human rights activists who take photos with their smartphones.
- Another app that operates in the same way and captures device sensor data is CameraV. It is an official tool of the New York-based organization that "helps people use video and technology to protect and defend human rights."
During our research, we encountered an article published on Medium, which included relevant information. While Medium's credibility can be doubted, because anyone can add their story on the website, the article came from Open Data Science, an established AI and machine learning-focused organization. Also, it provided multiple references to research papers, which we also attached.